AI risk research

1. Guarantees
2. Wiring
3. In this section

The optional counterparty-ai package adds qualitative, advisory risk context using an LLM with native tool use. It plugs into the same RiskStrategy seam as the rule-based default.

The AI never decides hard pass/fail. It consumes the finished report as ground truth, grounds every claim in a tool’s source URL (no source means inconclusive), and forces human review below a confidence threshold. Output is advisory only.

A separate package on purpose: prompts change often and must not bump the core’s version.

Guarantees

• Advisory only. Sanctions hits and VAT status stay deterministic in the report; the AI cannot override them.
• Grounded. Only findings with a source URL become Evidence; ungrounded claims are dropped to “needs review”.
• Human-in-the-loop. requiresHumanReview() is true on any adverse finding, any ungrounded claim, any inconclusive/adverse hard fact, or when confidence is below the threshold.
• Bounded cost. Results are PSR-16 cached by counterparty + report + prompt version.

Wiring

use Gawrys\Counterparty\Ai\AiRiskStrategy;
use Gawrys\Counterparty\Ai\Prompt\RiskPromptBuilder;
use Gawrys\Counterparty\Ai\Research\AnthropicResearchProvider; // or OpenAiResearchProvider
use Gawrys\Counterparty\Ai\Tool\{RegistryTool, ReportLookupTool};

$strategy = new AiRiskStrategy(
    provider: new AnthropicResearchProvider($http, $apiKey),
    promptBuilder: new RiskPromptBuilder(),
    tools: [new RegistryTool($registries), new ReportLookupTool()],
    cache: $psr16Cache,
    reviewThreshold: 0.6,
);

$verifier = new Verifier($checks, $strategy, $clock);

In this section

• Providers - Anthropic, OpenAI, and writing your own.
• Native tool use - how the model calls registry/report/web tools.
• Custom tools - expose your own data to the model.
• Testing - deterministic, offline tests with fakes and cassettes.

---

Table of contents

• Providers
• Native tool use
• Custom tools
• Testing